Skip to content

21 · Interview Preparation

Question sets with model answers for roles that build, design or secure systems on Claude and other large language models: AI engineer, solution architect and AI security engineer. Everything here is original; none of it comes from, or reconstructs, any certification exam.

Question sets

File Questions Best for
Conceptual questions 30 LLM fundamentals, prompting, context engineering, the API, tools, MCP, RAG, agents, evaluation
System design questions 26 Whiteboard prompts with clarifying questions, architecture, trade-offs and what interviewers look for
Security questions 28 Threat modelling, prompt injection, agent permissions, MCP, data protection, governance, incidents
Stakeholder and behavioural 26 Discovery, communicating trade-offs, delivery, leadership, ethics (STAR format)

Which sets for which role

Role Focus Also review
AI engineer Conceptual (all), system design 1–3, 9–10, 13–14, 20 Claude API cheat sheet, labs
Solution architect System design (all), stakeholder (all), conceptual 11–13, 20–30 13 · Enterprise AI architecture, CCAR-P one-pager
AI security engineer Security (all), system design 4, 11, 12, 15, 19, 25, conceptual 17–19 AI security cheat sheet, 12 · AI security
Technical leader Stakeholder (all), system design 6, 11, 12, 20, 26 13 · Business value and ROI

How to practise

  1. Answer out loud before reading the model answer. Interviews test spoken reasoning; reading answers silently builds false confidence.
  2. Time yourself. Conceptual answers in about two minutes; system design in 35–45 minutes; behavioural answers in about two minutes using STAR.
  3. Use your own experience. The behavioural model answers use fictional examples to show structure. Replace them with real situations you can discuss in depth. Never invent experience.
  4. Build something. Interviewers probe beyond the first answer. Having run the hands-on labs or a real-world project gives you concrete follow-up details.
  5. Keep facts current. Model names, prices and API features change. Answers here are dated 2026-10-24; check the cheat sheets and official docs before an interview.

What strong candidates do

Signal Example
Clarify before solving "How many documents, how often do they change, and do permissions vary per user?"
Prefer the simplest design "A routing workflow handles this; an agent adds cost without benefit here."
Measure "I'd build a 100-case eval set and compare cost per completed task."
Design for failure "Assume the model can be manipulated; enforce permissions in the tool."
Quantify "About 8,000 alerts a day, so Haiku for first-pass triage keeps cost manageable."
Communicate trade-offs "Option A is faster to ship; Option B is cheaper at scale. I recommend A for the pilot."

Common pitfalls

  • Jumping to multi-agent architectures for problems a single call can solve.
  • Describing prompt instructions as security controls.
  • Ignoring evaluation ("we'll test it manually").
  • Quoting outdated API patterns, such as temperature tuning or assistant prefill, which current Claude models reject.
  • Overclaiming accuracy or hiding limitations from stakeholders.